linux/drivers/oprofile/buffer_sync.c
<<
>>
Prefs
   1/**
   2 * @file buffer_sync.c
   3 *
   4 * @remark Copyright 2002-2009 OProfile authors
   5 * @remark Read the file COPYING
   6 *
   7 * @author John Levon <levon@movementarian.org>
   8 * @author Barry Kasindorf
   9 * @author Robert Richter <robert.richter@amd.com>
  10 *
  11 * This is the core of the buffer management. Each
  12 * CPU buffer is processed and entered into the
  13 * global event buffer. Such processing is necessary
  14 * in several circumstances, mentioned below.
  15 *
  16 * The processing does the job of converting the
  17 * transitory EIP value into a persistent dentry/offset
  18 * value that the profiler can record at its leisure.
  19 *
  20 * See fs/dcookies.c for a description of the dentry/offset
  21 * objects.
  22 */
  23
  24#include <linux/mm.h>
  25#include <linux/workqueue.h>
  26#include <linux/notifier.h>
  27#include <linux/dcookies.h>
  28#include <linux/profile.h>
  29#include <linux/module.h>
  30#include <linux/fs.h>
  31#include <linux/oprofile.h>
  32#include <linux/sched.h>
  33#include <linux/gfp.h>
  34
  35#include "oprofile_stats.h"
  36#include "event_buffer.h"
  37#include "cpu_buffer.h"
  38#include "buffer_sync.h"
  39
  40static LIST_HEAD(dying_tasks);
  41static LIST_HEAD(dead_tasks);
  42static cpumask_var_t marked_cpus;
  43static DEFINE_SPINLOCK(task_mortuary);
  44static void process_task_mortuary(void);
  45
  46/* Take ownership of the task struct and place it on the
  47 * list for processing. Only after two full buffer syncs
  48 * does the task eventually get freed, because by then
  49 * we are sure we will not reference it again.
  50 * Can be invoked from softirq via RCU callback due to
  51 * call_rcu() of the task struct, hence the _irqsave.
  52 */
  53static int
  54task_free_notify(struct notifier_block *self, unsigned long val, void *data)
  55{
  56        unsigned long flags;
  57        struct task_struct *task = data;
  58        spin_lock_irqsave(&task_mortuary, flags);
  59        list_add(&task->tasks, &dying_tasks);
  60        spin_unlock_irqrestore(&task_mortuary, flags);
  61        return NOTIFY_OK;
  62}
  63
  64
  65/* The task is on its way out. A sync of the buffer means we can catch
  66 * any remaining samples for this task.
  67 */
  68static int
  69task_exit_notify(struct notifier_block *self, unsigned long val, void *data)
  70{
  71        /* To avoid latency problems, we only process the current CPU,
  72         * hoping that most samples for the task are on this CPU
  73         */
  74        sync_buffer(raw_smp_processor_id());
  75        return 0;
  76}
  77
  78
  79/* The task is about to try a do_munmap(). We peek at what it's going to
  80 * do, and if it's an executable region, process the samples first, so
  81 * we don't lose any. This does not have to be exact, it's a QoI issue
  82 * only.
  83 */
  84static int
  85munmap_notify(struct notifier_block *self, unsigned long val, void *data)
  86{
  87        unsigned long addr = (unsigned long)data;
  88        struct mm_struct *mm = current->mm;
  89        struct vm_area_struct *mpnt;
  90
  91        down_read(&mm->mmap_sem);
  92
  93        mpnt = find_vma(mm, addr);
  94        if (mpnt && mpnt->vm_file && (mpnt->vm_flags & VM_EXEC)) {
  95                up_read(&mm->mmap_sem);
  96                /* To avoid latency problems, we only process the current CPU,
  97                 * hoping that most samples for the task are on this CPU
  98                 */
  99                sync_buffer(raw_smp_processor_id());
 100                return 0;
 101        }
 102
 103        up_read(&mm->mmap_sem);
 104        return 0;
 105}
 106
 107
 108/* We need to be told about new modules so we don't attribute to a previously
 109 * loaded module, or drop the samples on the floor.
 110 */
 111static int
 112module_load_notify(struct notifier_block *self, unsigned long val, void *data)
 113{
 114#ifdef CONFIG_MODULES
 115        if (val != MODULE_STATE_COMING)
 116                return 0;
 117
 118        /* FIXME: should we process all CPU buffers ? */
 119        mutex_lock(&buffer_mutex);
 120        add_event_entry(ESCAPE_CODE);
 121        add_event_entry(MODULE_LOADED_CODE);
 122        mutex_unlock(&buffer_mutex);
 123#endif
 124        return 0;
 125}
 126
 127
 128static struct notifier_block task_free_nb = {
 129        .notifier_call  = task_free_notify,
 130};
 131
 132static struct notifier_block task_exit_nb = {
 133        .notifier_call  = task_exit_notify,
 134};
 135
 136static struct notifier_block munmap_nb = {
 137        .notifier_call  = munmap_notify,
 138};
 139
 140static struct notifier_block module_load_nb = {
 141        .notifier_call = module_load_notify,
 142};
 143
 144static void free_all_tasks(void)
 145{
 146        /* make sure we don't leak task structs */
 147        process_task_mortuary();
 148        process_task_mortuary();
 149}
 150
 151int sync_start(void)
 152{
 153        int err;
 154
 155        if (!zalloc_cpumask_var(&marked_cpus, GFP_KERNEL))
 156                return -ENOMEM;
 157
 158        err = task_handoff_register(&task_free_nb);
 159        if (err)
 160                goto out1;
 161        err = profile_event_register(PROFILE_TASK_EXIT, &task_exit_nb);
 162        if (err)
 163                goto out2;
 164        err = profile_event_register(PROFILE_MUNMAP, &munmap_nb);
 165        if (err)
 166                goto out3;
 167        err = register_module_notifier(&module_load_nb);
 168        if (err)
 169                goto out4;
 170
 171        start_cpu_work();
 172
 173out:
 174        return err;
 175out4:
 176        profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
 177out3:
 178        profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
 179out2:
 180        task_handoff_unregister(&task_free_nb);
 181        free_all_tasks();
 182out1:
 183        free_cpumask_var(marked_cpus);
 184        goto out;
 185}
 186
 187
 188void sync_stop(void)
 189{
 190        end_cpu_work();
 191        unregister_module_notifier(&module_load_nb);
 192        profile_event_unregister(PROFILE_MUNMAP, &munmap_nb);
 193        profile_event_unregister(PROFILE_TASK_EXIT, &task_exit_nb);
 194        task_handoff_unregister(&task_free_nb);
 195        barrier();                      /* do all of the above first */
 196
 197        flush_cpu_work();
 198
 199        free_all_tasks();
 200        free_cpumask_var(marked_cpus);
 201}
 202
 203
 204/* Optimisation. We can manage without taking the dcookie sem
 205 * because we cannot reach this code without at least one
 206 * dcookie user still being registered (namely, the reader
 207 * of the event buffer). */
 208static inline unsigned long fast_get_dcookie(struct path *path)
 209{
 210        unsigned long cookie;
 211
 212        if (path->dentry->d_flags & DCACHE_COOKIE)
 213                return (unsigned long)path->dentry;
 214        get_dcookie(path, &cookie);
 215        return cookie;
 216}
 217
 218
 219/* Look up the dcookie for the task's mm->exe_file,
 220 * which corresponds loosely to "application name". This is
 221 * not strictly necessary but allows oprofile to associate
 222 * shared-library samples with particular applications
 223 */
 224static unsigned long get_exec_dcookie(struct mm_struct *mm)
 225{
 226        unsigned long cookie = NO_COOKIE;
 227
 228        if (mm && mm->exe_file)
 229                cookie = fast_get_dcookie(&mm->exe_file->f_path);
 230
 231        return cookie;
 232}
 233
 234
 235/* Convert the EIP value of a sample into a persistent dentry/offset
 236 * pair that can then be added to the global event buffer. We make
 237 * sure to do this lookup before a mm->mmap modification happens so
 238 * we don't lose track.
 239 */
 240static unsigned long
 241lookup_dcookie(struct mm_struct *mm, unsigned long addr, off_t *offset)
 242{
 243        unsigned long cookie = NO_COOKIE;
 244        struct vm_area_struct *vma;
 245
 246        for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
 247
 248                if (addr < vma->vm_start || addr >= vma->vm_end)
 249                        continue;
 250
 251                if (vma->vm_file) {
 252                        cookie = fast_get_dcookie(&vma->vm_file->f_path);
 253                        *offset = (vma->vm_pgoff << PAGE_SHIFT) + addr -
 254                                vma->vm_start;
 255                } else {
 256                        /* must be an anonymous map */
 257                        *offset = addr;
 258                }
 259
 260                break;
 261        }
 262
 263        if (!vma)
 264                cookie = INVALID_COOKIE;
 265
 266        return cookie;
 267}
 268
 269static unsigned long last_cookie = INVALID_COOKIE;
 270
 271static void add_cpu_switch(int i)
 272{
 273        add_event_entry(ESCAPE_CODE);
 274        add_event_entry(CPU_SWITCH_CODE);
 275        add_event_entry(i);
 276        last_cookie = INVALID_COOKIE;
 277}
 278
 279static void add_kernel_ctx_switch(unsigned int in_kernel)
 280{
 281        add_event_entry(ESCAPE_CODE);
 282        if (in_kernel)
 283                add_event_entry(KERNEL_ENTER_SWITCH_CODE);
 284        else
 285                add_event_entry(KERNEL_EXIT_SWITCH_CODE);
 286}
 287
 288static void
 289add_user_ctx_switch(struct task_struct const *task, unsigned long cookie)
 290{
 291        add_event_entry(ESCAPE_CODE);
 292        add_event_entry(CTX_SWITCH_CODE);
 293        add_event_entry(task->pid);
 294        add_event_entry(cookie);
 295        /* Another code for daemon back-compat */
 296        add_event_entry(ESCAPE_CODE);
 297        add_event_entry(CTX_TGID_CODE);
 298        add_event_entry(task->tgid);
 299}
 300
 301
 302static void add_cookie_switch(unsigned long cookie)
 303{
 304        add_event_entry(ESCAPE_CODE);
 305        add_event_entry(COOKIE_SWITCH_CODE);
 306        add_event_entry(cookie);
 307}
 308
 309
 310static void add_trace_begin(void)
 311{
 312        add_event_entry(ESCAPE_CODE);
 313        add_event_entry(TRACE_BEGIN_CODE);
 314}
 315
 316static void add_data(struct op_entry *entry, struct mm_struct *mm)
 317{
 318        unsigned long code, pc, val;
 319        unsigned long cookie;
 320        off_t offset;
 321
 322        if (!op_cpu_buffer_get_data(entry, &code))
 323                return;
 324        if (!op_cpu_buffer_get_data(entry, &pc))
 325                return;
 326        if (!op_cpu_buffer_get_size(entry))
 327                return;
 328
 329        if (mm) {
 330                cookie = lookup_dcookie(mm, pc, &offset);
 331
 332                if (cookie == NO_COOKIE)
 333                        offset = pc;
 334                if (cookie == INVALID_COOKIE) {
 335                        atomic_inc(&oprofile_stats.sample_lost_no_mapping);
 336                        offset = pc;
 337                }
 338                if (cookie != last_cookie) {
 339                        add_cookie_switch(cookie);
 340                        last_cookie = cookie;
 341                }
 342        } else
 343                offset = pc;
 344
 345        add_event_entry(ESCAPE_CODE);
 346        add_event_entry(code);
 347        add_event_entry(offset);        /* Offset from Dcookie */
 348
 349        while (op_cpu_buffer_get_data(entry, &val))
 350                add_event_entry(val);
 351}
 352
 353static inline void add_sample_entry(unsigned long offset, unsigned long event)
 354{
 355        add_event_entry(offset);
 356        add_event_entry(event);
 357}
 358
 359
 360/*
 361 * Add a sample to the global event buffer. If possible the
 362 * sample is converted into a persistent dentry/offset pair
 363 * for later lookup from userspace. Return 0 on failure.
 364 */
 365static int
 366add_sample(struct mm_struct *mm, struct op_sample *s, int in_kernel)
 367{
 368        unsigned long cookie;
 369        off_t offset;
 370
 371        if (in_kernel) {
 372                add_sample_entry(s->eip, s->event);
 373                return 1;
 374        }
 375
 376        /* add userspace sample */
 377
 378        if (!mm) {
 379                atomic_inc(&oprofile_stats.sample_lost_no_mm);
 380                return 0;
 381        }
 382
 383        cookie = lookup_dcookie(mm, s->eip, &offset);
 384
 385        if (cookie == INVALID_COOKIE) {
 386                atomic_inc(&oprofile_stats.sample_lost_no_mapping);
 387                return 0;
 388        }
 389
 390        if (cookie != last_cookie) {
 391                add_cookie_switch(cookie);
 392                last_cookie = cookie;
 393        }
 394
 395        add_sample_entry(offset, s->event);
 396
 397        return 1;
 398}
 399
 400
 401static void release_mm(struct mm_struct *mm)
 402{
 403        if (!mm)
 404                return;
 405        up_read(&mm->mmap_sem);
 406        mmput(mm);
 407}
 408
 409
 410static struct mm_struct *take_tasks_mm(struct task_struct *task)
 411{
 412        struct mm_struct *mm = get_task_mm(task);
 413        if (mm)
 414                down_read(&mm->mmap_sem);
 415        return mm;
 416}
 417
 418
 419static inline int is_code(unsigned long val)
 420{
 421        return val == ESCAPE_CODE;
 422}
 423
 424
 425/* Move tasks along towards death. Any tasks on dead_tasks
 426 * will definitely have no remaining references in any
 427 * CPU buffers at this point, because we use two lists,
 428 * and to have reached the list, it must have gone through
 429 * one full sync already.
 430 */
 431static void process_task_mortuary(void)
 432{
 433        unsigned long flags;
 434        LIST_HEAD(local_dead_tasks);
 435        struct task_struct *task;
 436        struct task_struct *ttask;
 437
 438        spin_lock_irqsave(&task_mortuary, flags);
 439
 440        list_splice_init(&dead_tasks, &local_dead_tasks);
 441        list_splice_init(&dying_tasks, &dead_tasks);
 442
 443        spin_unlock_irqrestore(&task_mortuary, flags);
 444
 445        list_for_each_entry_safe(task, ttask, &local_dead_tasks, tasks) {
 446                list_del(&task->tasks);
 447                free_task(task);
 448        }
 449}
 450
 451
 452static void mark_done(int cpu)
 453{
 454        int i;
 455
 456        cpumask_set_cpu(cpu, marked_cpus);
 457
 458        for_each_online_cpu(i) {
 459                if (!cpumask_test_cpu(i, marked_cpus))
 460                        return;
 461        }
 462
 463        /* All CPUs have been processed at least once,
 464         * we can process the mortuary once
 465         */
 466        process_task_mortuary();
 467
 468        cpumask_clear(marked_cpus);
 469}
 470
 471
 472/* FIXME: this is not sufficient if we implement syscall barrier backtrace
 473 * traversal, the code switch to sb_sample_start at first kernel enter/exit
 474 * switch so we need a fifth state and some special handling in sync_buffer()
 475 */
 476typedef enum {
 477        sb_bt_ignore = -2,
 478        sb_buffer_start,
 479        sb_bt_start,
 480        sb_sample_start,
 481} sync_buffer_state;
 482
 483/* Sync one of the CPU's buffers into the global event buffer.
 484 * Here we need to go through each batch of samples punctuated
 485 * by context switch notes, taking the task's mmap_sem and doing
 486 * lookup in task->mm->mmap to convert EIP into dcookie/offset
 487 * value.
 488 */
 489void sync_buffer(int cpu)
 490{
 491        struct mm_struct *mm = NULL;
 492        struct mm_struct *oldmm;
 493        unsigned long val;
 494        struct task_struct *new;
 495        unsigned long cookie = 0;
 496        int in_kernel = 1;
 497        sync_buffer_state state = sb_buffer_start;
 498        unsigned int i;
 499        unsigned long available;
 500        unsigned long flags;
 501        struct op_entry entry;
 502        struct op_sample *sample;
 503
 504        mutex_lock(&buffer_mutex);
 505
 506        add_cpu_switch(cpu);
 507
 508        op_cpu_buffer_reset(cpu);
 509        available = op_cpu_buffer_entries(cpu);
 510
 511        for (i = 0; i < available; ++i) {
 512                sample = op_cpu_buffer_read_entry(&entry, cpu);
 513                if (!sample)
 514                        break;
 515
 516                if (is_code(sample->eip)) {
 517                        flags = sample->event;
 518                        if (flags & TRACE_BEGIN) {
 519                                state = sb_bt_start;
 520                                add_trace_begin();
 521                        }
 522                        if (flags & KERNEL_CTX_SWITCH) {
 523                                /* kernel/userspace switch */
 524                                in_kernel = flags & IS_KERNEL;
 525                                if (state == sb_buffer_start)
 526                                        state = sb_sample_start;
 527                                add_kernel_ctx_switch(flags & IS_KERNEL);
 528                        }
 529                        if (flags & USER_CTX_SWITCH
 530                            && op_cpu_buffer_get_data(&entry, &val)) {
 531                                /* userspace context switch */
 532                                new = (struct task_struct *)val;
 533                                oldmm = mm;
 534                                release_mm(oldmm);
 535                                mm = take_tasks_mm(new);
 536                                if (mm != oldmm)
 537                                        cookie = get_exec_dcookie(mm);
 538                                add_user_ctx_switch(new, cookie);
 539                        }
 540                        if (op_cpu_buffer_get_size(&entry))
 541                                add_data(&entry, mm);
 542                        continue;
 543                }
 544
 545                if (state < sb_bt_start)
 546                        /* ignore sample */
 547                        continue;
 548
 549                if (add_sample(mm, sample, in_kernel))
 550                        continue;
 551
 552                /* ignore backtraces if failed to add a sample */
 553                if (state == sb_bt_start) {
 554                        state = sb_bt_ignore;
 555                        atomic_inc(&oprofile_stats.bt_lost_no_mapping);
 556                }
 557        }
 558        release_mm(mm);
 559
 560        mark_done(cpu);
 561
 562        mutex_unlock(&buffer_mutex);
 563}
 564
 565/* The function can be used to add a buffer worth of data directly to
 566 * the kernel buffer. The buffer is assumed to be a circular buffer.
 567 * Take the entries from index start and end at index end, wrapping
 568 * at max_entries.
 569 */
 570void oprofile_put_buff(unsigned long *buf, unsigned int start,
 571                       unsigned int stop, unsigned int max)
 572{
 573        int i;
 574
 575        i = start;
 576
 577        mutex_lock(&buffer_mutex);
 578        while (i != stop) {
 579                add_event_entry(buf[i++]);
 580
 581                if (i >= max)
 582                        i = 0;
 583        }
 584
 585        mutex_unlock(&buffer_mutex);
 586}
 587
 588
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.